Fuzzy Sets and Systems 130 (2002) 159–174

www.elsevier.com/locate/fss

Approximation theory of fuzzy systems based upon genuine

many-valued implications — MIMO cases 
Yong-Ming Lia; b; ∗ , Zhong-Ke Shib , Zhi-Hui Lia
a Department of Mathematics, Shaanxi Normal University, Xi’an 710062, People’s Republic of China
b Department of Automatic Control, Northwestern Polytechnical University, Xi’an 710072, People’s Republic of China

Received 30 December 1999; received in revised form 20 March 2001; accepted 3 April 2001

Abstract
It is constructively proved that the multi-input–multi-output fuzzy systems based upon genuine many-valued implications
are universal approximators (they are called Boolean type fuzzy systems in this paper). The general approach to construct
such fuzzy systems is given, that is, through the partition of the output region (by the given accuracy). Two examples are
provided to demonstrate the way in which fuzzy systems are designed to approximate given functions with a given required
approximation accuracy.
c 2002 Elsevier Science B.V. All rights reserved.

Keywords: Fuzzy system; Genuine many-valued implication; Universal approximator

1. Introduction nonlinear control solution with enough accuracy. For

In recent years, there have been a number of ap-
plications of fuzzy systems theory in various =elds,
for example, in control systems. In most of these ap-
plications, the main design objective is to construct a
fuzzy system to approximate a desired control or de-
cision (often experts). From a mathematical point of
view, fuzzy systems are just functions mapping their
input to output. In the context of control, the ques-
tion is whether a fuzzy controller can always be con-
structed to approximate any desired continuous and
 This work is supported by National Science Foundation of
China (Grant No. 19901028) and National 973 Foundational Re-
search Program G1998030417.
∗ Corresponding author. Fax: +86-29-530-7025.
E-mail address: liyongm@snnu.edu.cn (Y.-M. Li).
fuzzy systems used as models, the issue is whether a
fuzzy model can always be established which is ca-
pable of approximating any continuous and nonlinear
physical system arbitrarily well. The questions are of
both theoretical and practical importance. If the fuzzy
systems were proved to be universal approximators,
then one would feel more comfortable to utilize them
as controllers and models. If not, the fuzzy systems
should be used to solve only those control and mod-
eling problems that they are capable of.
Due to its importance, the issue of fuzzy systems
as universal approximators has drawn signi=cant at-
tention in the past few years and progress has been
made [1,5,6,11,13,15,17,19 – 24]. Consider a fuzzy
system that comprises four principle components:
fuzzi=er, fuzzy rule base, fuzzy inference engine, and

0165-0114/02/$ - see front matter  c 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 1 6 5 - 0 1 1 4 ( 0 1 ) 0 0 1 7 2 - 5
160 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174
defuzzi=er. Assume that the fuzzi=er is the most
commonly used singleton fuzzi=er, defuzzi=er is the
center of areas (COA) or averaging of maximums
(MOMs), and the fuzzy rule base for an multi-input–
single-output (MISO) system consists of rules in the
following forms:
(MISO) Rk : IF x1 is A1k AND x2 is A2k AND · · ·
AND xm is Am
k;
THEN y is Bk (k = 1; 2; : : : ; n);
(1)
where Akj (Bk ) in Uj (V ) about variable xk (y) are
linguistic terms characterized by fuzzy member-
ship function Akj (x j ) (Bk (y)). We generally con-
sider normal membership functions, such as trian-
gular, trapezoidal or Gaussian functions. Each Rk
can be viewed as a fuzzy implication (relation)
Ak = A1k × A2k × · · · × Am
k → Bk over U × V = (U1 ×
U2 × · · · × Um ) × V . The choice of fuzzy inference
is Mexible. This Mexibility is determined by the im-
plication operators chosen. For the previous approx-
imation work in [1,5,6,11,13,15,17,19 – 24], mainly
Mamdani type fuzzy systems and Takagi–Sugeno
(TS) type fuzzy systems, the implication operator
→ is always chosen as conjunctive type implica-
tion, that is, choosing → as a t-norm, such as min
(∧) and product operator. We call the corresponding
fuzzy systems the conjunctive type fuzzy systems
as in [6]. For the choice of implication operators as
genuine many-valued implications (i.e. the general-
ization of that of classic implications, mainly contain
S-implications, R-implications and QL-implications
[7–9]), at present, there do not exist any approxima-
tion results in the literature for such fuzzy systems.
Although J.L. Castro has done some work in this
respect as genuine many-valued implications chosen
as R-implications (with some restrictions), we shall
see later that his inference approach corresponding
to R-implications is not appropriate. However, as the
results of the experiment show [2 – 4,10], the Boolean
type fuzzy systems (we call the fuzzy system based
on genuine many-valued implications the Boolean
type fuzzy system as in [7]) have good control capa-
bility compared to other implication operators such as
conjunctive type implication operators. Furthermore,
since the fuzzy logic based on genuine many-valued
implications is a generalization of classic logic, it
has been studied as multi-valued logic systems [8,9]
and thus has a strict logic foundation [18], it also has
a widespread use in expert systems [2 – 4,10]. The
question “Are they also universal approximators
and how is their approximation mechanisms” still
remains unanswered. However, we have answered
this question for SISO Boolean type fuzzy systems
in [14]. In this paper, these results will be extended
to multi-input–multi-output (MIMO) fuzzy systems
based on genuine many-valued implications.
The structure of the rest of the paper is as follows.
In Section 2, we review the inference methods of
multi-rules fuzzy systems for MIMO fuzzy systems.
We discuss the approximation properties of MIMO
fuzzy systems based on R-implications in Section 3
and those of MIMO fuzzy systems based on S- and
QL-implications in Section 4. In Section 5, we shall
illustratively design two fuzzy systems based on gen-
uine many-valued implication for two given functions
with a required approximation accuracy and compare
the optimal fuzzy rules to those conjunctive fuzzy sys-
tems. Conclusions are made in the last section. One
appendix for the proofs of the propositions in the pa-
per is included.
2. The choice of multi-rules fuzzy inference
methods
We =rst review the (multi-rules) fuzzy inference
method presented in the previous paper [14]. Since the
MIMO fuzzy systems can always be separated into a
group of MISO fuzzy systems [12], without loss of
generality, we assume in this paper that fuzzy sys-
tems are MISO systems f : U ⊆ Rm → V ⊆ R as form
(1), where U = U1 × U2 × · · · × Um ⊆ Rm is the input
space and V ⊆ R is the output space. The MIMO ver-
sions of all results in this paper can be easily obtained
by doing a few simple manipulations.
For the rule Rk , its antecedent fuzzy set Ak = A1k ×
Ak × · · · ×Am
2
k on U = U1 × U2 × · · · ×Um with mem-
bership function Ak (x) = A1k (x1 ) ∗ A2k (x2 ) ∗ · · · ∗Am k
(xm ), where ∗ is the T-norm [8,9], x = (x1 ; x2 ; : : : ; xm ) ∈
U . Then each Rk can be viewed as fuzzy implica-
tion relation Ak = A1k × A2k × · · · × Amk → Bk , which
is a fuzzy set in U × V with a membership function
Rk (x; y) = Ak (x) → Bk (y).
 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174 161

For fuzzy system (1) and a given input A
, inference pseudo-triangle-shaped in the following, the discus-
approaches are presented as follows: sion for that of pseudo-trapezoid-shaped membership
m
function is similar.




B =A ◦

(Ak → Bk ); (2) We have the following general result.
k=1

m
Lemma 2.1 (Li et al. [14]). If the implication oper-




B =

A ◦ (Ak → Bk ); (3) ator is chosen as a t-norm; inference algorithm is as-
k=1
sumed as (4); then B
≡ 0.
m

B
= A
◦ (Ak → Bk ); (4) Hence, if we use inference method (4) and (5) for
k=1
conjunctive type implication operator, then the out-
put is always 0 for any input. In this case, the control
m
 is impossible for any processes; the inference method
B
= A
◦ (Ak → Bk ); (5) (2) and (3) should be used. This is also the theoret-
k=1
ical explanation of the related experiment results in
where ◦ denotes the relation composition operator, it is [3,4,10].
the just generalized Zadeh’s max-∗ composition rules, On the other hand, if → is chosen as genuine
where ∗ is a certain t-norm operator. many-valued implication, then I (0; x) ≡ 1 holds for
Assume that the input is a singleton x = x0 , then in any x ∈ [0; 1], in this case, we should use inference
the above algorithm, (2) and (3) are equivalent, and method (4) and (5) instead of (2) and (3), some
(4) and (5) are equivalent. reason is presented in the following lemma.
Assume that the requirement of the membership
function Ck (x) (x ∈ X ⊆ R), for each input variable Lemma 2.2 (Li et al. [14] and MSantaras [16]). If
and for each output variable in (1), is the same as the implication operator is chosen as a genuine many-
that of paper [23]. We assume that {Ck } is contin- valued implication; inference method is assumed as
uous, normal, consistent and complete, that is, for (2); then B
≡ 1.
any k, there exists x ∈ X such that Ck (x) = 1; and for
any x ∈ X , there exists k such that Ck (x)¿0; and if Remark 2.3. From the formal logic and matching
Ck (x) =1, then Cj (x) = 0 for any j=k. In particu- points of view, we can also give some explanation as
lar, if Ck (x) = 1 holds for any x ∈ X , then {Ck } is follows:
called a normal base set. For the background of this From the view of semantics of formal logic, fuzzy
requirement, we refer to [23]. We assume that Ck (x) inference based on genuine many-valued implications
is a pseudo-trapezoid-shaped membership function for assume that, if the truth-values of the antecedents of
any k, that is, Ck (x) has the following forms: the rules are false, then the truth-values of the rules are
true, and truth-values of the rules are non-increasing
Ck (x; xk ; bk ; ck ; xk+1 )
about the antecedents of the rules. However, in the

 M (x); x ∈ [xk ; bk ); actions of control, or in the common-sense inference,



 1; we always assume that, if the truth-values of the an-
x ∈ [bk ; ck );
= (6) tecedents of the rules are false or nearly false, then the

 D(x); x ∈ [ck ; xk+1 ]; corresponding rules shall not be =red or have little ef-


 fect on the =nal control. In this case, it is reasonable
0; x ∈ U − [xk ; xk+1 ];
to use the inference form (5) instead of (3) so as to
where xk 6bk 6ck 6xk+1 ; xk ¡xk+1 ; M (x)¿0 is diminish the eTect of those rules with small matching
strictly increasing in [xk ; bk ) and D(x)¿0 is strictly degree of the antecedents.
decreasing in [ck ; xk+1 ]. In particular, if bk = ck , then On the other hand, from the matching point of view,
Ck (x) is pseudo-triangle-shaped membership func- fuzzy inference based on conjunctive implications as-
tion. We assume that the membership function is sumes that, the bigger the matching degrees (or the
162 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174

truth-values) of the antecedents of the rules, the larger 3. Approximation capability of MISO fuzzy
is the contribution of the rules to the =nal control. So, systems based on R-implication
if the matching degrees of the antecedents of the rules
are zeros (false), then the corresponding rules shall not Suppose that I is an R-implication in this section,
be =red or have little aTect on the =nal control action. T is the corresponding t-norm. Then I : [0; 1] × [0; 1]
In this case, it is reasonable to use the inference form → [0; 1] is de=ned as follows:
(3) instead of (5).
I (a; b) = sup{c ∈ [0; 1] | T (a; c)6b}: (9)
Remark 2.4. From Remark 2.3, we give some notes The defuzzi=cation method is always assumed as the
about the constructions of the fuzzy systems based on averaging of MOA.
R-implications by Castro in [5]. For MISO fuzzy system (1), let x = x0 = (x01 ; : : : ;
For R-implications I : [0; 1] × [0; 1] → [0; 1] satis- m
x0 ) be a singleton input, then the fuzzy output based
fying the condition I (a; 0) = 0 if a=0, given input on I is
x = x0 , then Castro assumed that the output of the ith n

rules was B
(y) = (Ak (x0 ) → Bk (y)): (10)

 k=1
0; Ai (x0 ) = 0;
Bi
(y) = (7) Suppose that only Ak(1) (x0 ); : : : ; Ak(l) (x0 ) are not
I (Ai (x0 ); Bi (y)); Ai (x0 )=0 zero in the following, then Eq. (10) can be simply
rewritten as follows:
l

and the =nal fuzzy output was

B
(y) = (Ak(i) (x0 ) → Bk(i) (y)): (11)
i=1
B
(y) = Bi (y): (8)
i
Since we require {Bi } to be consistent, for any y ∈ V ,
there are at most two adjacent elements Bk ; Bk+1 such
that Bk (y)¿0; Bk+1 (y)¿0. In this case, Eq. (11) can
In fact, Castro used the inference method (3) instead
be calculated as follows:
of (5). As mentioned in the above remark, this is only

reasonable for the case I being a conjunctive impli- B
(y) = (Ak(i) (x0 ) → Bk (y)) ∧ (Ak(i) (x0 )
cation, so the choice of inference approach (8) using i∈T1 i∈T2
R-implications is not appropriate, even though there

has been some restriction on each consequent part of → Bk+1 (y)) (Ak(i) (x0 ) → 0): (12)
fuzzy rules, as (7). From the matching point of view, i∈T3
the required principle: for input x = x0 is the smaller
For input x = x0 , there is one determined output
the value (not necessary 0) Ai (x0 ), the smaller will be
y = y0 , for system (1). For this output y = y0 , it is
the consequent value Bi
(y) of the corresponding rule
reasonable to require that the value B
(y0 ) is mainly
Ri . However, choice (7) as a consequent value Bi
(y)
determined by the =rst two terms of Eq. (12), that is,
of the corresponding rule Ri is contrary to the required

principle. In fact, since I is an R-implication, I (a; b) is B
(y) = (Ak(i) (x0 ) → Bk (y)) ∧ (Ak(i) (x0 )
nonincreasing about the =rst variable a, given an input i∈T1 i∈T2
x = x0 , the smaller the value (not necessary 0) of the
→ Bk+1 (y)): (13)
antecedent Ai (x0 ), the larger is the value of the conse-
quent Bi
(y) (if Bi
(y) = 0) of the corresponding rule If we have the following assumption on fuzzy sys-
Ri . This forces B
(y), which is the supremum of all tem (1), then Eqs. (12) and (13) are equivalent.
Bi
(y), to be large (enough) at y as long as Bi
(y) = 0,
even though Ai (x0 ) is extremely small. This points out Assumption. For any singleton input x = x0 , if the
the inference approach (7) and (8) used by Castro in rules Rk(1) ; : : : ; Rk(l) are =red, that is, the member-
[5], for R-implications is not appropriate. ship values Ak(1) (x0 ); : : : ; Ak(l) (x0 ) of the antecedents
 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174 163

of these rules are not zero, then the membership func- Remark 3.3. (1) If f(z) = T (a; z) : [0; 1] → [0; a] is
tions Bk(1) (y); : : : ; Bk(l) (y) of the consequent part of strictly the increasing mapping for any a ∈ (0; 1], then
these rules have at most two adjacent di<erent ele- the solution of Eq. (15) is unique.
ments Bk (y); Bk+1 (y). (2) As for some examples, we give the solution of
(15) for some t-norms:
In fact, for some R-implications, this assumption is
necessary. (i) T = ∧; min operator: If a6b, then the solution of
Eq. (15) is
Lemma 3.1. If the membership functions {Bi } are

required to be consistent; and the implication op- 1=2; a ¿ 1=2;
erator I satis=es the condition I (a; 0) = 0 whenever z=
a = 0; then for the rule base (1); if we use infer- 1 − a; a61=2
ence method (10) as above; and the consequent part
Bk(1) (y); : : : ; Bk(l) (y) have at least 3 elements; then if a¿b, then the solution of (15) is
B
(y) ≡ 0.

In fact, the above assumption is also reasonable, 1=2; b ¿ 1=2;
z=
because many rules based on the practical fuzzy con- 1 − b; b 6 1=2
trol have the forms, and we can indeed construct such
kinds of fuzzy systems (in Theorem 3.9) below. This
(ii) T = •, product operator: z = b=(a + b).
requirement justi=es the following principle: If the in-
(iii) T (a; b) = max{0; a + b−1}, bounded sum:
puts or antecedent parts intersect; then the outputs
z = (1 + |b − a|)=2.
or consequent parts must intersect.
In this paper, we always make this assumption on
Corollary 3.4. For a; b ∈ (0; 1]; if a + b61; then
fuzzy systems. In this case, Eq. (13) always holds.
Eq. (15) has one solution z satisfying a61 − z and
For a Boolean implication →, since a → b is non-
b6z.
increasing to =rst variable a and T1 ; T2 are =nite sets of
{1; 2; 
: : : ; l} in Eq. (13), then there 
exists k1 ∈ T1 such
Theorem 3.5. Suppose that T is a continuous t-norm,
that i∈T1 (Ak(i) (x0 ) → Bk (y)) = ( i∈T1 Ak(i) (x0
)) →
I is the corresponding R-implication; then the system
Bk (y) = Ak1 (x0 ) → Bk (y) and k2 ∈ T2 such that i∈T2
function of MISO fuzzy systems (1) based on I is
(Ak(i) (x0 ) → Bk+1 (y)) = Ak2 (x0 ) → Bk+1 (y). Without
de=ned as follows: for any input x = x0 ; there are
loss of generality, we can assume that k1 = k; k2 = k +
two rules Rk ; Rk+1 such that Eq. (14) holds; then
1, then Eq. (13) can be simply rewritten as follows:

B
(y) = (Ak (x0 ) → Bk (y)) y0 = G(x0 ) = Bk−1 (z); (16)

∧(Ak+1 (x0 ) → Bk+1 (y)):



Let Y0 = {y0 | B (y0 ) = maxy∈V B (y)}, then the
control output y0 is one chosen point of Y0 (such as
the middle point of Y0 , but we do not make this restric-
tion). We thus obtain a function y0 = G(x0 ); x0 ∈ U ;
it is called the system function of (1) in this paper.
We =rst discuss the form of G(x) in the following
propositions.

Lemma 3.2. For any continuous t-norm T; the fol-
lowing equation has one solution for any a; b ∈ (0; 1]
(14)
where z is just the solution of Eq. (15) for
a = Ak+1 (x0 ); b = Ak (x0 ) and y0 = Bk−1 (z) satis=es
the condition Bk (y0 ) + Bk+1 (y0 ) = 1.
Corollary 3.6. If I is the Goguen implication; that
is; I (a; b) = 1 ∧ b=a; then the system function of (1)
based on I has the following form:
Ak (x0 )
y0 = G(x0 ) = Bk−1 :
Ak (x0 ) + Ak+1 (x0 )

T (a; z) = T (b; 1 − z): (15) This is from Theorem 3.5 and Remark 3.3(2)(ii).
164 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174

Corollary 3.7. If I is GAodel-implication; then the sys- ", there exists a natural number N such that
tem function of (1) based on I is as follows: (d − c)=N ¡"=2. Let e = (d − c)=N , suppose that
y1 = c; y2 = c + e; : : : ; yk+1 = c + ke; : : : ; yN +1 = d.
y0 = G(x0 ) Constructing membership functions B1 ; B2 ; : : : ; BN +1
 −1 such that the support of B1 is [y1 ; y2 ), the support of

 B (1=2); min{Ak (x0 ); Ak+1 (x0 )} BN +1 is (yN ; yN +1 ], the support of Bk is (yk−1 ; yk+1 )
 k


 ¿ 1=2; for 1¡k¡N + 1, and {Bk }Nk=1 +1
is a normal base set



 (such as the symmetric triangle-shaped membership
 B (1 − Ak+1 (x0 )); Ak+1 (x0 ) 6 1=2 and
−1
=
k functions), then Bk (yk ) = 1 for any k = 1; 2; : : : ; N +1.

 Ak+1 (x0 ) 6 Ak (x0 ); Let






−1
B (1 − Ak (x0 )); Ak (x0 ) 6 1=2 and

 k U1 = f−1 ([y1 ; y1 + 23 e)); : : : ; Uk

Ak+1 (x0 ) ¿ Ak (x0 ):
= f−1 ((yk − 23 e; yk + 23 e)); : : : ;
This is from Theorem 3.5 and Remark 3.3(2)(i).
UN +1 = f−1 ((yN +1 − 23 e; yN +1 ])
Corollary 3.8. If I is Lukasitwicz implication; that then {U1 ; U2 ; : : : ; UN +1 } forms an open cover of U
is; I (a; b) = (1 − a + b) ∧ 1; then the system function and Ui1 ∩ Ui2 = ∅ if and only if i1 ; i2 are the adjacent
of (1) has the following form: numbers.
For the given e, since f is uniform continuous
y0 = G(x0 ) = Bk−1 (1=2 + |Ak (x0 ) − Ak+1 (x0 )|=2): over compact set U , there exists a positive num-
ber %, such that |f(x) − f(x
)|¡e=2 whenever
This is from Theorem 3.5 and Remark 3.3(2)(iii). d(x; x
)¡%, where d(x; x
) = max{|x1 −x1
|; |y1 −y1
|},
For fuzzy systems based on R-implication opera- x = (x1 ; y1 ), x
= (x1
; y1
). Therefore, there exists a =-
tors, the results in [3,4,10] showed their good con- nite set F = {x1 = (x11 ; x21 ); : : : ; xs = (x1s ; x2s )} such that
trol capability compared to other implication operators the %=2-neighborhood of F forms another open cover
such as conjunctive type implication operators. Then, of U; that is, {(x1i − %=2; x1i + %=2) × (x2i − %=2; x2i +
whether or not the fuzzy systems based upon these im- %=2) | i = 1; : : : ; s} forms an open cover of U , and the
plications are universal approximators, the following following appropriate condition holds.
theorem answers this problem.
Appropriate condition. Write the partition points of
Theorem 3.9. The fuzzy systems based on R- [a1 ; b1 ] and [a2 ; b2 ] formed by F by x11 ¡ · · · ¡xl1 ; x12
implications with MOA defuzzi=er are universal ¡ · · · ¡xd2 (then |xi1 − xi+11
|¡% and |xi2 − xi+12
|¡%);
approximators. That is; for any continuous func- 1 1 2 2
then for any open rectangle (xi ; xi+2 ) × (xj ; xj+2 ); it
tion f : U → R over a compact subset U ⊆ Rm and meets {Uk } at most two adjacent elements.
an arbitrary given positive number "; there is a
fuzzy system; its corresponding system function Constructing the normal base set {A1i } for [a1 ; b1 ]
y = G(x) (as given in Theorem 3:5) based on a given and {A2i } for [a2 ; b2 ] such that the support of A1k is
1 1
R-implication with MOA defuzzi=er satis=es the (xk−1 ; xk+1 ) and the support of A2k is (xk−1
2 2
; xk+1 ), then
1 1 2 2
inequality relation Ak (xk ) = 1 and Ak (xk ) = 1 for any k. Fuzzy rules base
is designed as follows:
max |f(x) − G(x)|¡":
x∈U
Rij : IF x1 is A1i AND x2 is A2j ;
Proof. For simplicity and better presentation, we will THEN y Cij: (i = 1; 2; : : : ; l; j = 1; 2; : : : ; d));
prove the case of two variables (i.e., m = 2). The proof
for more variables is similar. where Cij is chosen as there exists k such that xi ∈ Uk
Since f is continuous, the image of f is also (there are at most two Uk such that xi ∈ Uk as the
a closed interval assumed as [c; d]. For the given construction of Uk ), then Cij is chosen as Bk .
 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174 165

As the appropriate condition and the properties of As the assumption for the fuzzy rule base in this pa-
the open cover {Uk }, the assumption of this paper for per, there are three cases for consequent membership
the rule base holds. This is because, for any single- functions:
ton input x = x0 ; if the rules Rk(1) ; : : : ; Rk(l) are =red, (1) Bi; j (y) = Bi+1; j (y) = Bi; j+1 (y) = Bi+1; j+1 (y) =
that is, the membership values Ak(1) (x0 ); : : : ; Ak(l) (x0 ) Bk . In this case, since Bk is symmetrically triangle-
of the antecedents of these rules are not zero, then shaped and Bk (yk ) = 1, then there exists non-negative
the membership functions Bk(1) (y); : : : ; Bk(l) (y) of the real number % such that for any y ∈ [yk − %; yk + %];
consequent part of these rules have at most two adja- B
(y) attains its maximum value 1. It follows that
cent diTerent elements Bk ; Bk+1 .
y0 = G(x0 ) = yk : (17a)
For any input x0 ∈ [a1 ; b1 ] × [a2 ; b2 ], if x0 ∈ [xi1 ; xi+1
1
]
× [xj2 ; xj+1
2
], then (2) Bi; j (y) = Bi+1; j (y) = Bi; j+1 (y) = Bk ; Bi+1; j+1 (y)

= Bk+1 . Let a = A1i (x01 ), b = A2j (x02 ), as Bk ; Bk+1 are
B (y) = Ai (x0 ) → Ci (y) ∧ Ai+1 (x0 ) → Ci+1 (y):
triangle-shaped, it follows that
Suppose that Ci = Bk , then Ci+1 =Bk+1 ; Bk or Bk−1 as y − yk+1 y − yk+1
the appropriate condition. In this case, y0 = Ci−1 (Ai B
(y) = ∧
ab(yk − yk+1 ) a(1 − b)(yk − yk+1 )
(x0 )) (as in Theorem 3.5) is in [yk−1 ; yk ] or [yk ; yk+1 ].
It is no hurt to assume that y0 ∈ [yk ; yk+1 ], then y − yk+1
∧
max{|y0 −yk |; |y0 −yk+1 |}6e. Noting that x0 belongs (1 − a)b(yk − yk+1 )
to at most two adjacent open sets Uk or Uk+1 , it follows y − yk
that |f(x0 ) − f(yk )|¡ 23 e or |f(x0 ) − f(yk+1 )|¡ 23 e; ∧ ∧1
(1 − a)(1 − b)(yk+1 − yk )
and thus
y − yk+1
=
|G(x0 ) − f(x0 )| max{ab; a(1 − b); b(1 − a)}(yk − yk+1 )
= |y0 − f(x0 )| y − yk
∧ ∧ 1:
(1 − a)(1 − b)(yk+1 − yk )
6 min{|y0 − yk | + |yk − f(x0 )|; |y0 − yk+1 |
There are three cases
+|yk+1 − f(x0 )|} (2a) b¿a and a61=2. In this case,
6 e + 23 e = 53 e ¡ 56 " ¡ ": y − yk+1
B
(y) =
(1 − a)b(yk − yk+1 )
Hence, maxx∈U {|G(x) − f(x)|}¡".
Furthermore, if all the consequent membership y − yk
∧ ∧1
functions {Bi } are assumed as symmetric triangle- (1 − a)(1 − b)(yk+1 − yk )
shaped, then the system function y = G(x) de=ned by and y0 = G(x0 ) satis=es the following equation:
(16) has the simple form that we calculated in the fol-
lowing. As an example, only two inputs fuzzy systems y0 − yk+1
are considered. R-implication is chosen as Goguen- (1 − a)b(yk − yk+1 )
implication, t-norm is just a product operation. y 0 − yk
In this case, for any singleton input x0 = (x01 ; x02 ), =
(1 − a)(1 − b)(yk+1 − yk )
there are at most four rules =red, and Eq. (10) has the
following form: ⇒ (1 − b)y0 − (1 − b)yk+1 = −by0 + byk

B
(y) = A1i (x01 )A2j (x02 ) → Bij (y) ⇒ y0 = byk + (1 − b)yk+1 : (17b)
In this case, the four =red rules Ri; j ; Ri+1; j ; Ri; j+1 ;
∧ A1i (x01 )A2j+1 (x02 ) → Bi; j+1 (y)
Ri+1; j+1 reduce to two simple rules
∧ A1i+1 (x01 )A2j (x02 ) → Bi+1; j (y) IF x2 is A2j THEN y is Bk ;
∧ A1i+1 (x01 )A2j+1 (x02 ) → Bi+1; j+1 (y): IF x2 is A2j+1 THEN y is Bk+1 :
166 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174

(2b) b¿0:5 and a¿0:5. In this case, y 0 − yk

= ⇒ (1 − a)(y0 − yk+1 )
b(1 − a)(yk+1 − yk )
y − yk+1
B
(y) = = −a(y0 − yk ) ⇒ y0 = ayk + (1 − a)yk+1 :
ab(yk − yk+1 ) (17e)
y − yk
∧ ∧1
(1 − a)(1 − b)(yk+1 − yk )
(3b) b60:5. In this case, y0 = G(x0 ) satis=es the
and y0 = G(x0 ) satis=es the following equation: following equation:
y0 − yk+1
y0 − yk+1 y0 − yk
= a(1 − b)(yk − yk+1 )
ab(yk − yk+1 ) (1 − a)(1 − b)(yk+1 − yk )
y 0 − yk
⇒ (1 − a)(1 − b)(y0 − yk+1 ) =
(1 − a)(1 − b)(yk+1 − yk )
= −ab(y0 − yk ) ⇒ y0
⇒ (1 − a)(y0 − yk+1 )
1 − a − b + ab
= yk = −a(y0 − yk ) ⇒ y0 = ayk + (1 − a)yk+1 :
1 − a − b + 2ab
(17f)
ab
+ yk+1 : (17c)
1 − a − b + 2ab
As in cases (2b) and (2c), the four =red rules are
reduced to two simple rules in case (3).
(2c) a¿b and a60:5. In this case, as in case (2a),
The combination of formulas (17a) – (17f) is
we can obtain that
just the analytic representation of systems function
y0 = ayk + (1 − a)yk+1 : (17d) y = G(x1 ; x2 ) for this kind of fuzzy system. Example
2 in Section 5 illustrates the formula.
The four =red rules were reduced to two simple In fact, the above discussion to two-input fuzzy
rules as in case (2b). systems (1) reMects certain decoupling properties of
Case (2) can be seen as three of the four =red conse- fuzzy systems based on R-implications, and this kind
quent membership functions that are same, while the of properties should be researched in the future.
other is diTerent.
(3) Bi; j (y) = Bi+1; j (y) =Bk ; Bi; j+1 (y) =Bi+1; j+1 (y) Remark 3.10. The results of Theorem 3.9 show the
= Bk+1 , two of four =red consequent membership approximation capability of MISO Fuzzy systems
functions are the same and the other two are the other based on R-implications, and the proof of Theo-
similar ones. In this case, rem 3.9 gives an approach to explicitly construct
such a fuzzy system, in particular, its consequent
y − yk+1 part is constructive. It is not like other approaches
B
(y) =
max{ab; a(1 − b)}(yk − yk+1 ) [1,5,6,11,13,15,17,19 – 24], where their membership
y − yk functions in the consequent part cover each other
∧ not satisfying the consistent conditions or consequent
max{b(1 − a); (1 − a)(1 − b)}(yk+1 − yk )
part that is simply chosen as a real number. We
∧1: actually provide an approach to construct normal,
complete and consistent membership functions in the
There are two cases: consequent part.
(3a) b¿0:5. In this case, y0 = G(x0 ) satis=es the The approximation mechanism of MISO fuzzy
following equation: systems based on R-implications is similar to that
of conjunctive type implications, that is, the nearer
y0 − yk+1 the distances of the input x = x0 and the antecedent
ab(yk − yk+1 ) membership Ai , the nearer the distances of the output
 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174
y0 = Bk−1 (z) and the corresponding consequent mem-
bership Bi will be, and the output y0 is only related
with the 2m adjacent base elements of the input x = x0 .

4. Approximation capability of MISO
fuzzy systems based on S-implications and
QL-implications
S-implication and QL-implication are de=ned, re-
spectively as follows:
I (a; b) = S(N (a); b); I (a; b) = S(N (a); T (a; b));(18)

where S is a t-conorm, T is a t-norm, N is a negation,
S and T are dual through N [8 – 10]. They both gener-
alize classical genuine many-valued logic and satisfy
the following condition
I (0; b) = 1: (19)
They generally do not satisfy the condition I (a; b) = 1
whenever a6b, so the discussion of Section 3 does
not hold for S- and QL-implications. We shall give
other analyses in the following.
The requirements of membership functions are the
same as those of Theorem 3.9. In this case, inference
formula (14) also holds, that is,
B
(y) = S(N (Ai (x0 )); Bi (y))
∧S(N (Ai+1 (x0 )); Bi+1 (y)); (20)
B
(y) = S(N (Ai (x0 )); T (Ai (x0 ); Bi (y)))
∧S(N (Ai+1 (x0 )); T (Ai (x0 ); Bi+1 (y))): (21)
The defuzzi=er is also assumed as MOA.
Lemma 4.1. Let I (a; b) = (1 − a) ∨ b be Kleene–
Dienes implication; I (a; b) = (1 − a) ∨ (a ∧ b) be
Zadeh implication; then the system function corre-
sponding to fuzzy system (1) is
 planation of related experiment results in [3,4,7,10].
 y; Ai (x0 ) ¿ Ai+1 (x0 );
 i
y = G(x) = yi+1 ; Ai (x0 ) ¡ Ai+1 (x0 );


(yi + yi+1 )=2; Ai (x0 ) = Ai+1 (x0 );
(22)
which lies in the interval [yi ; yi+1 ]; the intersection
supports of Bi and Bi+1 . Generally; G(x) is not con-
tinuous.
167
Lemma 4.2. For S-implication I (a; b) = 1 − a + ab;
Reichenbach implication; the system function of
fuzzy system (1) is
−1 Ai (x0 )
y = Bi ; (23)
Ai (x0 ) + Ai+1 (x0 )
which lies in the interval [yi ; yi+1 ]; the intersection
supports of Bi and Bi+1 .
Lemma 4.3. For QL-implication I (a; b) = 1−a+a2 b
and fuzzy system (1); its system function is
Ai (x0 ) − Ai+1 (x0 ) + A2i+1 (x)
y = Bi−1 ; (24)
A2i (x) + A2i+1 (x)
which lies in the interval [yi ; yi+1 ]; the intersection
supports of Bi and Bi+1 .
Although the corresponding system functions of
fuzzy system (1) based on diTerent S-implications and
QL-implications are not the same, the following ap-
proximation theorem always holds:
Theorem 4.4. MISO fuzzy systems based on
S-implications and QL-implications with MOA de-
fuzzi=er are also universal approximators.
Remark 4.5. Since the formulas of fuzzy systems
based on diTerent kinds of S-implications and QL-
implications are diTerent, the corresponding system
functions of fuzzy system (1) are not the same.
Therefore, their corresponding control capabilities
are diTerent. Furthermore, from the results of Lem-
mas 4.2 and 4.3, we can implement continuous
control action with diTerent weights through the S-
implication I (a; b) = 1 − a + ab and QL-implication
I (a; b) = 1 − a + a2 b. In this case, we have a better
control capability. These results give a theoretic ex-
5. Examples
In the following examples, we shall use our ap-
proach to approximate the given continuous functions.
The comparison of optimal fuzzy rules by our ap-
proach and those by Ying in [21] is presented.
168 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174
0.9
B1 B2 B3
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0.1 0.2 0.3 0.4
Fig. 1. The membership functions of the output variable in Example 1.
Example 1. Design a fuzzy system based on genuine
many-valued implication to approximate the continu-
ous function F(x) = sin (x)=x de=ned on [−3; 3] with
" = 0:2.
The image set of F(x) on [−3; 3] is [0:0470; 1].
Design the fuzzy system through dividing the image
set [0:0470; 1] into [(1 − 0:0470)=0:2] + 1 = 5 parts,
and then decide the input fuzzy membership functions.
The membership functions of the consequent part
are B1 ; B2 ; B3 ; B4 ; B5 ; B6 , which are depicted as follows
(Fig. 1).
The membership functions of input variables are
A1 ; A2 ; : : : ; A11 , which are depicted as follows (Fig. 2).
Fuzzy rule base is
R1 : IF x is A1 or A11 ; THEN y is B1 ;
R2 : IF x is A2 or A10 ; THEN y is B2 ;
R3 : IF x is A3 or A9 ; THEN y is B3 ;
R4 : IF x is A4 or A8 ; THEN y is B4 ;
R5 : IF x is A5 or A7 ; THEN y is B5 ;
R6 : IF x is A6 ; THEN y is B6 :
Using R-implication as implication operator, the
systems function of this fuzzy system is depicted in the
following graph (Fig. 3). A comparison of the graph
of F(x) = sin x=x is also depicted in the same =gure.
The approximation error is " = 0:0864¡0:2.
In this example, we use only 6 rules to approximate
F(x) = sin x=x with an accuracy " = 0:2, while Ying in
[21] used 207 rules to approximate the same function.
B4 B5 B6
0.5 0.6 0.7 0.8 0.9 1
Example 2. Design a fuzzy system based on genuine
many-valued implication to uniformly approximate
the polynomial z = P(x1 ; x2 ) = 0:52 + 0:1x1 + 0:38x2 −
0:06x1 x2 de=ned on [−1; 1]2 with " = 0:1.
The image set of P over [−1; 1]2 is [−0:02; 0:94].
Design the fuzzy system by dividing the image set
[−0:02; 0:94] into [(0:94 − (−0:02))=0:1] + 1 = 10
parts, and then decide the input fuzzy membership
functions.
The membership functions of the consequent part
are B1 ; B2 ; : : : ; B11 , which are depicted as follows
(Fig. 4).
The membership functions of input variables x1
and x2 are A11 ; A12 ; : : : ; A17 and A21 ; A22 ; : : : ; A219 , they are
depicted as follows (Figs. 5 and 6), respectively.
Table 1 gives us Fuzzy rule base as follows:
In this table, the natural number 1; 2; 3; : : : represents
the index number of the membership function for the
related variable x1 ; x2 or z.
Using Goguen implication, the =gure of system
function G(x1 ; x2 ) is depicted as follows (Fig. 7). A
comparison of the =gure of P(x1 ; x2 ) and the =gure of
the errors of the system function (or approximation
function) and the origin function P(x1 ; x2 ) are also de-
picted in Figs. 8 and 9.
The approximation error is " = 0:084¡0:1.
In this example, we use only 133 rules (not consid-
ering decoupling) to approximate P(x1 ; x2 ) = 0:52 +
0:1x1 + 0:38x2 − 0:06x1 x2 with an accuracy " = 0:1,
while Ying in [21] used 225 rules to approximate the
same function.
 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174 169

0.9
A1 A2 A3 A4 ...... A8 A9 A10 A11
0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
-3 -2 -1 0 1 2 3

Fig. 2. The membership functions of the input variable in Example 1.

0.9

0.8

0.7

0.6

y 0.5

0.4

0.3

0.2

0.1

0
-3 -2 -1 0 1 2 3
x

Fig. 3. The comparison of system function y = G(x) (solid-line) and the origin function y = F(x) (dotted-line).

0.9
B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 B11
0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

Fig. 4. The membership functions of the output variable in Example 2.

170 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174

0.9
A1 A2 A3 A4 A5 A6 A7
0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1

Fig. 5. The membership functions of the =rst input variable in Example 2, where Ai represents A1i (i = 1; : : : ; 7).

0.9
A1 A2 A3 A4 ... A16 A17 A18 A19
0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1

Fig. 6. The membership functions of the second input variable in Example 2, where Ai represents A2i (i = 1; : : : ; 19).

Table 1
Z values

x2

x1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

1 1 2 2 3 3 4 4 5 5 6 6 7 7 8 8 9 9 10 10
2 2 2 2 3 3 4 4 5 5 6 6 7 7 8 8 9 9 10 10
3 2 3 3 3 3 4 5 5 6 6 7 7 8 8 9 9 9 10 11
4 3 3 3 4 4 4 5 6 6 7 7 8 8 9 9 9 10 10 11
5 4 4 4 4 5 5 5 6 7 7 8 8 8 9 9 9 10 10 11
6 4 4 5 5 5 6 6 6 7 7 8 8 9 9 9 10 10 11 11
7 5 5 5 5 6 6 6 7 7 7 8 8 9 9 9 10 10 11 11

6. Conclusion results:

We discuss the approximation theory of MIMO (1) The fuzzy systems based upon R-, S-, QL-
Boolean type fuzzy systems and obtain the following implications are universal approximators, their
 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174
Fig. 7. The system function z = G(x1 ; x2 ) of the fuzzy system in
Example 2 based on Goguen-implication.
Fig. 8. The =gure of origin function z = P(x1 ; x2 ) in Example 2.
corresponding formulas of system functions are
given.
(2) General approaches to construct fuzzy sys-
tems are given; two examples illustrate the
approaches.
(3) Defuzzi=cation methods and inference methods
are given for Boolean type fuzzy systems.
The optimal fuzzy rules and decoupling properties
of MIMO (i.e. (1)) Boolean type fuzzy systems should
be further researched in the future.
171
Fig. 9. Errors of the origin function and approximation function
(system function) in Example 2.
Appendix
Proof of Lemma 3.1. As the choice of implication I
and {Bi }, it follows that, if Bk(1) (y); : : : ; Bk(l) (y) have
at least three elements, then for any y ∈ V , there exists
i(y); 16i(y)6l such that Bk(i) (y) = 0. In this case,
Ak(i) (x0 ) = 0 and Ak(i) (x0 ) → Bk(i) (y) = Ak(i) (x0 ) → 0
l
= 0. It follows that B
(y) = i = 1 (Ak(i) (x0 ) → Bk(i)
(y)) = 0. Hence, B
(y) ≡ 0.
Proof of Lemma 3.2. Since T is a continuous t-norm,
the functions f : [0; 1] → [0; a] and g : [0; 1] → [0; b]
de=ned by f(z) = T (a; z); g(z) = T (b; 1 − z) are
continuous surjections, f(z) is nonincreasing and
g(z) is nondecreasing. Let h(z) = f(z) − g(z), then
h(0) = f(0) − g(0) = T (a; 0) − T (b; 1) = −b and
h(1) = f(1) − g(1) = T (a; 1) − T (b; 0) = a, and thus
h : [0; 1] → [−b; a] is a nonincreasing continuous sur-
jection. Since a; b¿0 and h(0) = −b60; h(1) = a¿0,
there is at least one point z ∈ [0; 1] such that h(z) = 0.
This z is just the solution of Eq. (15).
Proof of Corollary 3.4. Since a + b61, it follows
that a61 − b; b61 − a, then [b; 1 − a] = ∅. For
any z ∈ [b; 1 − a], h(z) = T (a; z) − T (b; 1 − z) has
the following properties: h(b) = T (a; b) − T (b; 1 −
z) = T (b; a) − T (b; 1 − b)60 and h(1 − a) = T (a; 1 −
a) − T (b; a) = T (a; 1 − a) − T (a; b)¿0. Since h(z) is
continuous, it follows that there exists one z ∈ [b; 1−a]
172 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174
such that h(z) = 0, this is just to say that Eq. (15) has
one solution z satisfying a61 − z and b6z.
Proof of Theorem 3.5. For any input x = x0 , there
are two rules Rk ; Rk+1 to make Eq. (14) hold, that is,
B
(y)=(Ak (x0 )→Bk (y))∧(Ak+1 (x0 )→Bk+1 (y)). Let
F(x0 ; y) = B
(y), since →; T and ∧ are continuous,
F(x0 ; y) is continuous with respect to x0 and y.
Thus, for any x0 , there exists one y0 ∈ V such that
B
(y0 ) = maxy∈V B
(y), and thus y0 = G(x0 ) is also
continuous w.r.t. x0 .
Since Ak (x0 ) → Bk (y) and Ak+1 (x0 ) → Bk+1 (y) are
both continuous membership functions with nonempty
intersections, so the y that B
(y) gets its maximum
value from satis=es the following equation:
(Ak (x0 ) → Bk (y)) = (Ak+1 (x0 ) → Bk+1 (y)): (A.1)
Let Y1 = {y ∈ V | Ak (x0 )6Bk (y) and Ak+1 (x0 )6
Bk+1 (y)}, if Y1 = ∅, then Ak (x0 ) + Ak+1 (x0 )6Bk (y)
+ Bk+1 (y) = 1, from Corollary 3:4, there exists
a solution z of Eq. (15) such that b = Ak (x0 )6z
and a = Ak+1 (x0 )61 − z. Let z = Bk (y) and 1 −
z = Bk+1 (y), since Bk (y) is pseudo-triangle-shaped,
there exists one unique point y0 = Bk−1 (z) satisfying
Bk (y0 ) + Bk+1 (y0 ) = 1 and y0 ∈ Y1 .
If Y1 = ∅, then we must have Ak (x0 )¿Bk (y) and
Ak+1 (x0 )¿Bk+1 (y). Since T is continuous, for any
a ∈ (0; 1], the map T (a; ) : [0; 1] → [0; a] is always
surjective, then we should have
T (Ak (x0 ); Ak (x0 ) → Bk (y)) = Bk (y);
T (Ak+1 (x0 ); Ak+1 (x0 ) → Bk+1 (y)) = Bk+1 (y): (A.2)
From Eqs. (A:1) and (A:2), we can obtain the fol-
lowing equation:
T (Ak+1 (x0 ); Bk (y)) = T (Ak (x0 ); Bk+1 (y)): (A.3)
Let z = Bk (y), then z is just the solution of Eq. (15)
for a = Ak+1 (x0 ); b = Ak (x0 ). In this case, (A:3) has
only one solution y0 = Bk−1 (z) satisfying Bk (y0 ) +
Bk+1 (y0 ) = 1.
Proof of (Appropriate Condition) of Theorem 3.9.
Otherwise, there exist at least three elements
Uj1 ; Uj2 ; Uj3 that intersect with E = (xi1 ; xi+2
1
) × xj2 ;
2
xj+2 ). Without loss of generality, it is no hurt to
assume that j1 ¡j2 ¡j3 , then Uj1 ∩ Uj3 = ∅. Choos-
ing al = (a1l ; a2l ) ∈ Ujl ∩ (xi ; xi+2 ) (l = 1; 2; 3), then it
follows that
f(al ) ∈ (yjl − 23 e; yjl + 23 e)
and thus, |f(a1 ) − f(a3 )|¿(yj3 − 23 e) − (yj1 + 23 e)
= (yj3 − yj1 ) − 43 e. Being the choice of yj3 ; yj1 , it fol-
lows that yj3 − yj1 ¿2e; and thus,
|f(a1 ) − f(a3 )| ¿ 2e − 43 e = 23 e ¿ 12 e: (A.4)
1
On the other hand, since a1 ; a3 ∈ E and d((xi+2 ;
2 1 2 1 2 1 2 1
xj+2 ); (xi ; xj ))6d((xi+2 ; xj+2 ); (xi+1 ; xj+1 )) + d((xi+1 ;
2
xj+1 ); (xi1 ; xj2 ))6%=2 + %=2 = %, it follows that |a1
− a3 | ¡%, and thus, |f(a1 ) − f(a3 )| ¡ e=2, this
inequality contradicts with the inequality (A.4). This
proves that the appropriate condition holds.
To prove Lemma 4:1, we need the following lemma.
Lemma A. For two di<erent membership functions
Ai = A1i ∗ · · · ∗ Am 1 m
i and Aj = Aj ∗ · · · ∗ Aj in MISO
fuzzy system (1); where ∗ is any t-norm; for any input
x0 ∈ U = U1 × · · · × Um ; we always have Ai (x0 ) +
Aj (x0 )61.
Proof. If Ai (x0 ) = 0 or Aj (x0 ) = 0, then Ai (x0 ) +
Aj (x0 )61 is obvious.
If Ai (x0 ) = 0 or Aj (x0 ) = 0, then for all 16i6m, Aki
(x0 ) = 0 and Ajk (x0 ) = 0. Since {Aki } is a norm base
for variable xk ∈ Uk , there is at least one k such that
Aki (x0 ) + Ajk (x0 ) = 1, then it follows that Ai (x0 ) +
Aj (x0 ) = A1i ∗· · ·∗Am 1 m k
i (x0 )+Aj ∗· · ·∗Aj (x0 )6Ai (x0 )+
k
Aj (x0 ) = 1.
Proof of Lemma 4.1. In this case, Eqs. (20) and (21)
can be rewritten as follows:
B
(y) = [(1 − Ai (x0 )) ∨ Bi (y)]
∧[(1 − Ai+1 (x0 )) ∨ Bi+1 (y)]; (A.5)
B
(y) = [(1 − Ai (x0 )) ∨ (Ai (x0 ) ∧ Bi (y))]
∧[(1 − Ai+1 (x0 )) ∨ (Ai+1 (x0 ) ∧ Bi+1 (y))];
(A.6)
 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174 173

where x0 ∈ (xi ; xi+1 ). Write a = Ai (x0 ); b = Ai+1 (x0 ) Proof of Lemma 4.2. In this case, Eq. (22) can be
and z = Bi (y), since a + b = Ai (x0 ) + Ai+1 (x0 )61 rewritten as follows:
(Lemma A), Eqs. (A:5) and (A:6) are as
follows: B
(y) = [1 − Ai (x0 ) + Ai (x0 )Bi (y)]

B
(y) = [1 − (a ∨ b)] ∨ [(1 − b) ∧ z] ∧[1 − Ai+1 (x0 ) + Ai+1 (x0 )Bi+1 (y)]: (A.11)

∨[(1 − a) ∧ (1 − z)] ∨ [z ∧ (1 − z)];
B
(y) = [1 − (a ∨ b)] ∨ [(1 − b) ∧ a ∧ z]
∨[(1 − a) ∧ b ∧ (1 − z)]
∨[a ∧ b ∧ z ∧ (1 − z)]
then B
(y) can be calculated as follows:
B
(y)
(A.7) It is no hurt to assume that the support set of
Bi (y) and Bi+1 (y) is (yi−1 ; yi+1 ) and (yi ; yi+2 ), re-
spectively, in the following. We declare that the
unique point that B
(y) gets its maximum value
lies in [yi ; yi+1 ]. First, if y ∈= [yi−1 ; yi+2 ], then
B
(y) = [1 − Ai (x0 )] ∧ [1 − Ai+1 (x0 )] = 1 − (a ∨ b),
where Ai (x0 ) = a; Ai+1 (x0 ) = b, it follows that if
(A.8) there exists y such that B
(y)¿1 − (a ∨ b), then
y ∈ (yi−1 ; yi+2 ). We prove that there exists y such that
B
(y)¿1 − (a ∨ b) in the following. Let z = Bi (y),
then

 B
(y) = (1 − a + az) ∧ [1 − b + b(1 − z)]

 (1 − a) ∨ [(1 − b) ∧ z]; a ¿ b;


= (1 − b) ∨ [(1 − a) ∧ (1 − z)]; a ¡ b; = (1 − a + az) ∧ (1 − bz): (A.12)



1 − a; a = b; In order to demand B
(y)¿1 − (a ∨ b), it suUces to
(A.9) require that the following inequality has a solution z
for any a ∈ [0; 1]

B
(y) 1 − a + az ¿ 1 − (a ∨ b);

 (1 − a) ∨ [a ∧ z]; a ¿ b; 1 − bz ¿ 1 − (a ∨ b): (A.13)


= (1 − b) ∨ [b ∧ (1 − z)]; a ¡ b; The above inequality can be rewritten as follows:



1 − a; a = b; a(1 − z) ¡ a ∨ b;
(A.10)
bz ¡ a ∨ b: (A.14)


Since Bi (y) is symmetric, the point that B (y) gets its
The above equality holds for any 0¡z¡1.
maximum value for the case a¿b must satisfy the re-
For a =xed a ∈ [0; 1], because z1 = 1 − a + az and
lation Bi+1 (y)¿1−b¿a, in this case y0 = yi ; and sim-
z2 = 1−bz represent two direct lines, its slope is a and
ilarly, y0 = yi+1 for the case a¡b; for the case a = b,
(−b), respectively, then z1 is strictly increasing about
the maximum value 1−a of B
(y) gets at all y ∈ [c; d],
z and z2 is strictly decreasing about z. The point that
in this case, we obtain that y0 = (c + d)=2, this value
B
(y) gets its maximum value is the intersection point
will aTect the approximation capability of the corre-
of these two direct lines, then it satis=es the equation
sponding fuzzy system, so we make some corrections
z = a=(a + b), that is
in this case: as the =red consequent membership func-
tions are only Bi and Bi+1 , then it is reasonable that Ai (x0 )
y0 lies in (yi ; yi+1 ) (we assume that the support set Bi (y0 ) = and
Ai (x0 ) + Ai+1 (x0 )
of Bi (y) and Bi+1 (y) is (yi−1 ; yi+1 ) and (yi ; yi+2 ), re-
spectively), then we get y0 = (yi + yi+1 )=2. This com- Ai+1 (x0 )
Bi+1 (y0 ) = : (A.15)
pletes the proof. Ai (x0 ) + Ai+1 (x0 )
174 Y.-M. Li et al. / Fuzzy Sets and Systems 130 (2002) 159–174
For the input x0 ∈ [xi ; xi+1 ], Bi (y) and Bi+1 (y) are the
only two =red consequent parts, then y0 must lie in
the intersections of the supports of Bi (y) and Bi+1 (y),
that is, y0 ∈ [yi ; yi+1 ]. Since Bi (y) is strictly mono-
tone in the interval [yi ; yi+1 ], there exists a unique
point
 [11] B. Kosko, Fuzzy systems as universal approximators, IEEE
−1 Ai (x0 )
y0 = Bi
Ai (x0 ) + Ai+1 (x0 )
to make Eq. (A:15) hold, this is just the output of the
fuzzy system.
The proof of Lemma 4:3 is similar to that of
Lemma 4:2.
The proof of Theorem 4:4 is similar to that of
Theorem 3:9.
References
[1] J.J. Buckley, Universal fuzzy controllers, Automatica 18 (6)
(1992) 1245–1248.
[2] A.J. BugarSVn, S. Barro, Reasoning with truth values on
compacted fuzzy chained rules, IEEE Trans. Systems, Man
Cybernet. — Part B 28 (1) (1998) 34–46.
[3] Z. Cao, A. Kandel, Applicability of some fuzzy implication
operators, Fuzzy Sets and Systems 31 (1989) 151–186.
[4] Z. Cao, D. Park, A. Kandel, Investigations on the applicability
of fuzzy inference, Fuzzy Sets and Systems 49 (1992) 151–
169.
[5] J.L. Castro, Fuzzy logic controllers are universal
approximators, IEEE Trans. Systems, Man Cybernet. 25 (4)
(1995) 629–635.
[6] J.L. Castro, M. Delgado, Fuzzy systems with defuzzi=cation
are universal approximators, IEEE Trans. Systems, Man
Cybernet. — Part B 26 (1) (1996) 149–152.
[7] O. CordSon, F. Herrera, A. PeregrSVn, Applicability of the fuzzy
operators in the design of fuzzy logic controllers, Fuzzy Sets
and Systems 86 (1997) 15–41.
[8] D. Dubois, H. Prade, Fuzzy sets in approximate reasoning,
Part 1: inference with possibility distribution, Fuzzy Sets and
Systems 40 (1991) 143–202.
[9] D. Dubois, J. Lang, H. Prade, Fuzzy sets in approximate
reasoning, Part 2: logical approaches, Fuzzy Sets and Systems
40 (1991) 203–244.
[10] J.B. Kiszka, M.E. Kochanska, D.S. Sliwinska, The inMuence
of some fuzzy implication operators on the accuracy of a
fuzzy model — Part 1, Part 2. Fuzzy Sets and Systems 15
(1985) 111–128, 223–240.
Trans. Comput. 43 (11) (1994) 1329–1333.
[12] C.C. Lee, Fuzzy logic in control systems: fuzzy logic
controller, IEEE Trans. Systems Man Cybernet. 20 (1990)
404–435.
[13] Y.M. Li, Z.K. Shi, Approximation capability of fuzzy I=O
systems based on similarity measures, WCICA’2000, Hefei,
China, June 28 – July 2, 2000, pp. 1790–1794.
[14] Y.M. Li, Z.K. Shi, Z.H. Li, Approximation theory
of fuzzy systems based upon genuine many-valued
implications—SISO cases, Fuzzy Sets and Systems, in this
issue.
[15] D. Lissin, M.A. Gennert, Proceedings NAFIPS 1999, pp.
184 –188.
[16] L. MSantaras, Approximate reasoning models, Ellis Horwood
Series in Arti=cial Intelligence, Wiley, 1990, 68p.
[17] H.T. Nguyen, V. Kreinovich, O. Sirisaugtakin, Fuzzy control
as a universal control tool, Fuzzy Sets and Systems 80 (1996)
71–86.
[18] G.J. Wang, On the logic foundation of fuzzy reasoning,
Inform. Sci. 117 (1) (1999) 47–88.
[19] L.X. Wang, J.M. Mendel, Fuzzy basic functions, universal
approximation, and orthogonal least-square learning, IEEE
Trans. Neural Networks 3 (5) (1992) 807–814.
[20] P.Z. Wang, S.H. Tan, F.M. Song, P. Liang, Constructive
theory for fuzzy systems, Fuzzy Sets and Systems 88 (1997)
195–203.
[21] H. Ying, SuUcient conditions on general fuzzy systems as
function approximators, Automatic 30 (3) (1994) 521–525.
[22] H. Ying, SuUcient conditions on uniform approximation of
multivariate functions by Takagi–Sugeno fuzzy systems with
linear rule consequent, IEEE Trans. Systems, Man Cybernet.
— Part A 28 (4) (1998) 515–520.
[23] X.J. Zeng, M.G. Singh, Approximation theory of fuzzy
systems — SISO case, IEEE Trans Fuzzy Systems 2 (2)
(1994) 162–176.
[24] X.J. Zeng, M.G. Singh, Approximation theory of fuzzy
systems — MIMO case, IEEE Trans Fuzzy Systems 3 (2)
(1995) 219–235.